Title: Computation-Aware Kalman Filtering with Model Selection for Neural Dynamics

Abstract:

Bayesian techniques have long been central to the dynamical latent variable modeling of single-cell neural recordings, largely owing to their capacity to incorporate explicit priors and quantify uncertainty. Nevertheless, the proliferation of modern-scale datasets has shifted preference toward overparameterized deep networks, which offer superior predictive performance and more favorable computational scaling. Although numerous posterior approximation methods are available, each introduces some degree of approximation error. While recent studies have begun to address this issue by incorporating computational uncertainty, such approaches typically suffer from quadratic complexity and rely on fixed model hyperparameters.

In this work, we expand upon these developments by integrating model selection, supported by a novel training loss function and optimization strategy, thereby enabling tractable inference even in large state-spaces. We propose the Computation-Aware State-Space Model (CASSM), a framework tailored specifically for the scale-imbalanced regime, defined by scenarios where the number of trials is substantially lower than the count of recorded neurons. Through experiments on both synthetic and real-world data, we demonstrate that our approach remains competitive with data-intensive deep networks while offering significantly better uncertainty calibration than prior efforts to scale Bayesian methods. Ultimately, our findings provide neuroscience researchers with a practical guide for selecting among various dynamical latent variable models based on specific dataset characteristics and constraints.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC